Prediction of Peak Ground Velocity (PGV) and Cumulative Absolute Velocity (CAV) of Earthquakes Using Machine Learning Techniques

This study presents the prediction of cumulative absolute velocity (CAV) and peak ground velocity (PGV) using machine learning (ML) algorithms, which are relatively new compared to ground motion models with fixed functional forms. The performance of three ML algorithms, namely Linear Regression, Artificial Neural Network, and Gradient Boosting are evaluated and compared. The New Turkish Strong Motion Database (N-TSMD), containing over 23,000 recordings of 743 earthquakes that occurred in Turkiye between 1983 and 2020, is used to build ML models. In addition to N-TSMD, new recordings, including the recent M_w 7.7 and M_w 7.6 (Kahramanmaraş), M_w 6.6 (Gaziantep), and M_w 6.4 (Hatay) earthquakes, are added. In developing ML models, the moment magnitude (M_w), Joyner-Boore distance (R_JB), shear-wave velocity averaged in the top 30 m of soil (Vs₃₀), and style-of-faulting (SoF) are used as estimator parameters to characterize the source, path, site, and tectonic environment. Mean square error (MSE), root mean squared error (RMSE), and correlation coefficient (R) metrics are used to evaluate models. Results indicated that the Gradient Boosting algorithm demonstrates the best performance in predicting CAV and PGV according to all performance metrics. This is followed by Artificial Neural Network and Linear Regression, respectively. Residual analyses with predictions of the Gradient Boosting model indicated that there is almost no trend in the distribution of the total residuals of both PGV and CAV. The GB model’s prediction skill can be considered fair in all M_w, R_JB, and Vs₃₀ ranges.